Automata Learning for Behavioural Mapping of Zigbee Devices

Abstract

With the rise of the Internet of Things (IoT) in recent years, extra care needs to be taken to ensure that connected cyber-physical systems are secure. This thesis investigates the potential of active Automata Learning to model Zigbee devices. Zigbee is a radio protocol utilized in home automation, industrial, and commercial applications. Models help to understand the internal behaviour of black box systems. Active Automata Learning is a Machine Learning application to create a model. Resulting models can be used to prove certain properties, such as the security of a system, adherence to specified requirements (correctness), or to generate test cases for the system under test.
In this work we establish a model learning application over a radio link, addressing the growing need to model wireless communication protocols effectively. We have focussed on the tools available for model creation, the aspects of Zigbee that can be modelled as finite automata, and the impact of input choices and abstraction levels on model accuracy.
Using a mixed-methods approach that includes a literature review, design and creation of a testbed, and experimental validation, the research establishes a workflow for model learning of Zigbee. The findings demonstrate that specific abstraction levels and input configurations can influence the resulting models and an automated and structured testing environment is crucial for effective model learning. This work contributes to the field by providing insights into the modelling of complex radio protocols and lays the groundwork for future research in Automata Learning applications within IoT systems, opening the door to other applications such as fuzzing using the established testbed.

Date of Award	16 May 2025
Original language	English
Supervisor	Joshua Moerman (Examiner) & Fabian van den Broek (Co-assessor)